368 - Trajectory-based Respiratory Models Improve Prediction of Bronchopulmonary Dysplasia in Preterm Infants

Publication Number: 1352.368

Christopher Myers, Yale School of Medicine, New Haven, CT, United States; Waldemar Carlo, University of Alabama, Birmingham, AL, United States; Namasivayam Ambalavanan, University of Alabama School of Medicine, Birmingham, AL, United States; Samuel Gentle, Yale School of Medicine, New Haven, CT, United States

Background: Bronchopulmonary dysplasia (BPD) remains the most frequent morbidity in preterm infants. Traditional prediction models use baseline characteristics and respiratory support on specific postnatal days, but patient-specific trajectories in respiratory support during the first postnatal month may improve risk stratification.

Objective: Our objective is to compare the predictive accuracy of trajectory-based models versus traditional single timepoint logistic regression models for BPD prediction in preterm infants.

Design/Methods: This is a retrospective cohort study in preterm infants born at < 29 weeks' gestation with respiratory support assessments on postnatal days 1, 7, 14, and 28. Latent class growth analysis (LCGA) identified distinct trajectory patterns in respiratory support. Linear and quadratic trajectory models with 1-5 classes were tested according to Bayesian Information Criterion, entropy, and clinical interpretability. We then compared the ability of several logistic regressions to predict grade 2 or 3 BPD or death using Receiver Operating Characteristic curves. These included baseline demographics (birth weight, gestational age, sex, and antenatal steroids) without trajectory classes, demographics with respiratory support and fraction of inspired oxygen (FiO2) on postnatal days 1, 7, 14, and 28, and demographics with our trajectory classes. Trajectory model accuracy improvement compared to each time-point was assessed using exact prediction probability comparisons. Confidence intervals calculated using proportion tests.

Results: Among the 872 preterm infants, the three-class quadratic model fit our data best (BIC=10,886.6, entropy=0.829) , while identifying clinically distinct trajectories: Class 1 (moderate-stable n=253, 29%), Class 2 (moderate-improving, n=256, 29%), and Class 3 (severe-stable, n=363, 42%). Rates of grade 2 or 3 BPD or death varied significantly by class (p < 0.001): 13.4%, 5.1%, 53.2%, respectively. The trajectory model demonstrated superior accuracy compared to single timepoint models across postnatal days 1, 14, and 28, but not day 7 (p < 0.01 for all). At day 1, 57% of infants showed improved accuracy with the trajectory model (95% CI: 53-61%), at day 14 (63%, 95% CI: 59-67%) and day 28 (64%, 95% CI: 60-68%). The model did not improve accuracy when compared to day 7 (39%, 95% CI: 35-43%).

Conclusion(s): Trajectory-based prediction models provide superior accuracy compared to single-timepoint assessments for BPD prediction. The benefit is most pronounced when incorporating mid-to-late time periods, with nearly two-thirds of infants showing improved prediction at days 14 and 28.

Figure 1: Respiratory Support Trajectories and Individual Patient Patterns in Preterm Infants.
The three-class Quadratic Trajectory model demonstrated the lowest Bayesian Inference Criteria (BIC = 10886.6) of tested models, while maintaining a satisfactory entropy (0.829) and at least 50 subjects in each class. Each class is shown with the individual subjects that contribute to its respective trajectory.

Table 1: Trajectory Class Characteristics


Figure 2: Accuracy Improvement of Trajectory-Based Model Over Single-Timepoint for Predicting Grade 2-3 BPD or Death
Proportion of infants for whom the trajectory model demonstrated improved accuracy compared to single-timepoint models at Days 1, 7, 14, and 28. Error bars represent 95% confidence intervals. The red dashed line indicates no improvement (50%). The trajectory model showed superior accuracy at Day 1 (57%), Day 14 (63%), and Day 28 (64%), and reduced accuracy at Day 7 (39%, p<0.01).